Products

What We Build

We build open-source technology for the AI and HPC era, from a full datacenter you can spin up in software to the training that grows the engineers who run it. Dreamed up and engineered in Ireland, shared openly with the world. Work that is still in the laboratory is on the Projects page.

Cnuas Virtual AI HPC Infrastructure

A complete AI/HPC datacenter, in software. Cnuas is a 100% software-emulated infrastructure, with virtual GPUs, virtual NVLink-equivalent fabric, virtual RoCE and InfiniBand TOR switches and PCIe-attached virtual NICs, running unmodified Linux software stacks inside guests. Most of it also runs directly on the host, with no hypervisor, guest or kernel module.

Anyone with a laptop can stand up a multi-node GPU cluster, exercise the same libibverbs and CUDA-equivalent APIs used in production, and learn or prototype against a real-feeling fabric, without owning a single InfiniBand cable.

  • CnuasGPU, virtual GPU PCIe device with SIMT execution model
  • CnuasLink, GPU-to-GPU peer fabric (NVLink-equivalent)
  • CnuasSwitch, virtual TOR switch, RoCEv2 and InfiniBand multiplexed
  • CnuasNIC, virtual RoCE-IB NIC kernel driver and verbs provider
  • cnuas, unified CLI and REST API to configure and drive the whole system
  • Cnuas Compute, CUDA-equivalent GPU runtime (libcnuasrt.so) with scalar, AVX2 and AVX-512 backends
  • cnuassmi, nvidia-smi-equivalent GPU inventory and telemetry
  • CnuasBLAS v0.1, dense linear algebra (libcnuasblas.so), a cuBLAS equivalent over the runtime kernels
  • In development, the cnuascc compiler, CnuasCCL and CnuasSHMEM collectives, and the CnuasDNN/FFT/SPARSE/SOLVER math libraries
  • Standard Linux Verbs API, so guests use libibverbs unmodified
  • Rack architecture based on the Open Compute Project Open Rack v3 specifications
  • Facility Twin, an early-stage experimental extension exploring facility modeling with OpenUSD and Isaac Sim
Explore Cnuas Documentation Datasheets Research and availability
🖥️

CnuasGPU with CnuasLink

SIMT virtual GPU and the inter-GPU fabric that connects them. The full NVIDIA-equivalent compute stack, with a CUDA Runtime equivalent (Cnuas Compute) and an nvidia-smi equivalent (cnuassmi) shipping today, and a cnuascc compiler that lowers device source to CnuasIR for the scalar subset, with vector code generation in development.

🔀

CnuasSwitch with CnuasNIC

Top-of-rack switch daemon and PCIe-attached virtual NIC. Handles RoCEv2 over UDP and full InfiniBand including LRH routing, with DCB (PFC, ETS and ECN) modelled end to end.

📚

Used by PacketFive Academy

Cnuas is the lab environment for PacketFive Academy courses, where students run real distributed-training workloads against the emulated fabric, no hardware required.

PacketFive Academy
Academic Courses & Training Programme
Education CUDA ROCm RDMA OFED MPI

PacketFive Academy Academic Courses & Training Programme

High-performance Computing, AI and Networking, PacketFive's open-source training programme for the full AI/HPC software stack. Delivered through the PacketFive Academy learning platform, PacketFive Academy gives students, engineers and practitioners the deep technical skills modern AI clusters demand.

  • Structured courses on CUDA, ROCm, RDMA, RoCE, OFED, Open MPI and InfiniBand
  • PacketFive Academy, a full LMS with progress tracking and certificates of completion
  • Practical labs designed around the Cnuas Virtual AI HPC Infrastructure (no expensive hardware needed)
  • Instructor and self-paced learning tracks
  • Google SSO, JWT auth, and email verification built-in
  • Open-source content and platform under permissive licences
PacketFive Academy ↗ PacketFive Academy ↗

Open Source Commitment

All PacketFive R&D output, including code, training material, papers and reference implementations, is published on GitHub under permissive open-source licences.

PacketFive on GitHub ↗